Passive Nonlinear Observer Design for Ships Using. Lyapunov Methods: Full-Scale Experiments with a. Supply Vessel

Passive Nonlinear Observer Design for Ships Using Lyapunov Methods: Full-Scale Experiments with a Supply Vessel Thor I. Fossen a;? Jann Peter Strand a a Department of Engineering Cybernetics, Norwegian University of Science and Technology, N-74 Trondheim, Norway, e-mail: tif@itk.ntnu.no, jann@itk.ntnu.no Dynamic positioning (DP) and tracking systems for ships are usually designed under the assumption that the kinematic equations can be linearized about a set of predened constant yaw angles, typically 6 operating points in steps of degrees, to cover the whole heading envelope. This is necessary when applying linear (Kalman lter) theory and gain scheduling techniques. However, global exponential stability (GES) cannot be guaranteed if linear theory is used. In this paper a nonlinear observer is derived. The observer is proven to be passive and GES. The number of tuning parameters is reduced to a minimum by using passivity theory. This results in a simple and intuitive tuning procedure. The proposed observer includes features like estimation of both the low-frequency position and velocity of the ship from noisy position measurements, bias state estimation (environmental disturbances) and wave ltering. The nonlinear passive observer has been simulated on a computer model of a supply vessel and implemented on full-scale ships with excellent results. Copyright c998 Elsevier Science Ltd. Key words: Nonlinear observers, passive elements, marine systems, ship control. Introduction Dynamic positioning (DP) systems have been commercially available for marine vessels since the 96's. The rst DP systems were designed using conven-? This work was partially supported by ABB Industri, Marine Division and the Norwegian Research Council. Preprint submitted to Elsevier Preprint 7 May 998

tional PID controllers in cascade with low pass and/or notch lters to suppress the wave-induced motion components. From the middle of the 97's more advanced control techniques based on optimal control and Kalman-lter theory were proposed by Balchen et al. (976). This work has later been modied and extended by Balchen et al. (98a, 98b), Grimble et al. (98a, 98b), Fung and Grimble (98), Slid et al. (98) and more lately by Fossen et al. (996) and Srensen et al. (996). Filtering and state estimation are important features of a DP system. In most cases, measurements of the vessel velocities are not available. Hence, estimates of the velocities must be computed from noisy position and heading measurements through a state observer. Unfortunately, the position and heading measurements are corrupted with colored noise due to wind, waves and ocean currents as well as sensor noise. However, only the slowly-varying disturbances should be counteracted by the propulsion system, whereas the oscillatoric motion due to the waves (st-order wave disturbances) should not enter the feedback loop. This is done by using so-called wave ltering techniques, which separates the position and heading measurements into a low-frequency (LF) and wave frequency (WF) position and heading estimate (Fossen 994). In existing DP systems the wave ltering and state estimation problem are solved by using linear Kalman lters. The major drawback of this approach is that the kinematic equations of motions must be linearized about a set of predened constant yaw angles, typically 6 operating points in steps of degrees, to cover the whole heading envelope. For each of these linearized models, optimal Kalman lter and feedback control gains are computed such that the gains can be modied on-line by using gain-scheduling techniques. In the seek for new control strategies the linear Kalman lter approach is a major obstacle since it is dicult and time-consuming to tune the state estimator (stochastic system with 5 states and covariance equations). The main reason for this, is that the numerous covariance tuning parameters may be dicult to relate to physical quantities resulting in a somewhat ad hoc tuning procedure. In this paper a nonlinear passive observer is proposed. The observer includes wave ltering properties, bias state estimation, reconstruction of the LF motion components and noise-free estimates of the non-measured vessel velocities. The proposed observer is proven to be passive and globally exponentially stable (GES). Hence, only one set of observer gains are needed to cover the whole state space. In addition, the number of observer tuning parameters are signicantly reduced and the wave lter parameters are directly coupled to the dominating wave frequency. Passivity theory showed to be a new tool with respect to accurate tuning of the observer. The proposed nonlinear observer opens for new controller designs more in line with the actual structure of the physical system e.g. by using the separation principle (Fossen, Loria and

Panteley, 998) or observer backstepping. In Fossen and Grvlen (998) the concept of vectorial observer backstepping has been applied to derive a GES output feedback control system for ships. This work is, however, based on a simplied model of the environmental disturbances since it is assumed that the WF motion and bias states can be neglected in the design. More recently, Aarset, Strand and Fossen (998) have shown that these results can be extended to the general case by including a dynamic model for wave ltering and bias state estimation. The main difference between the observer backstepping design and the nonlinear observer presented in this paper is that passivity is not guaranteed in the observer backstepping case. We believe that this is an important feature since passivity arguments will simplify the tuning procedure signicantly. Hence, the time needed for sea-trials and tuning will be reduced. The performance and robustness of the observer are demonstrated by computer simulations and full-scale experiments of a thruster controlled supply vessel in the North Sea. The simulation study veries that all estimation errors converge exponentially to zero. The nonlinear passive observer, as opposed to a linearized Kalman-lter, guarantees exponential convergence of all bias estimation errors to zero. This is almost impossible to obtain in commercial DP systems using linearized Kalman-lters. Mathematical Modelling of Ships. Kinematic Equations of Motion Let the earth-xed position (x; y) and heading of the vessel relative to an earth-xed frame X E Y E Z E be expressed in vector form by = [x; y; ] T ; and let the velocities decomposed in a vessel-xed reference frame be represented by the state vector = [u; v; r] T. These three modes are referred to as the surge, sway and yaw modes of a ship. The origin of the vessel-xed frame XY Z is located at the vessel centre line in a distance x G from the center of gravity, see Figure. The transformation between the vessel-xed and the earth-xed velocity vectors is: _ = J() () where J() is a state dependent transformation matrix. Conventional ships are not equipped with actuators in roll and pitch which suggests that the roll and pitch modes should be omitted when designing output feedback controllers for ships. In fact, this is an appropriate assumption since both the rolling and

y X X E ( x, y ) Y Y E Fig.. Denition of the earth-xed X E Y E Z E and the vessel-xed XY Z reference frames. pitching motions of a ship is oscillatoric with zero mean and limited amplitude. Moreover, a conventional ship is metacentric stable which implies that there exist restoring moments in roll and pitch. In the remaining of the paper, it is assumed that the ship is sucient metacentric stable such that only the rotation matrix in yaw can be used to describe the kinematic equations of motion, that is: J() = J( ) = 64 cos? sin sin cos () 75 where J( ) is non-singular for all : Notice that J? ( ) = J T ( ).. Ship Modelling The LF motion of a large class of surface ships can be described by the following model (Fossen 994): M _ + D = + J T ()b () = B u u (4) Here < is a control vector of forces and moment provided by the propulsion system, that is main propellers aft of the ship and thrusters which can produce surge and sway forces as well as a yaw moment. In addition to this, ships can be equipped with control surfaces and rudders. The control inputs are denoted by u < r (r ) and B u < r is a constant matrix describing the actuator conguration. Unmodeled external forces and moment due 4

to wind, currents and waves are lumped together into an earth-xed constant (or slowly-varying) bias term b <. If a small Froude number is assumed, the inertia matrix M < which includes hydrodynamic added inertia can be written (Fossen 994): M = 64 m? X _u m? Y _v mx G? Y _r (5) 75 mx G? N _v I z? N _r where m is the vessel mass and I z is the moment of inertia about the vessel- xed z-axis. For control applications which are restricted to low-frequency motions, wave frequency independence of added inertia (zero wave frequency) can be assumed. This implies that _M =. The zero-frequency added mass in surge, sway and yaw due to accelerations along the corresponding axes are dened as X _u <, Y _v < and N _r < ; whereas the cross-terms Y _r and N _v can have both signs (SNAME 95). For a straight-line stable ship, D < will be a strictly positive damping matrix due to linear wave drift damping and laminar skin friction. The linear damping matrix is dened as (Fossen 994): D = 64?X u?y v mu? Y r?n v mx G u? N r 75 : (6) where the cruise-speed u = in DP and u > when moving forward. In general the damping forces will be nonlinear. However, for DP and cruising at constant speed linear damping is a good assumption since this a regulation problem (Fossen 994).. Bias Modelling (Slowly-Varying Environmental Disturbances) It is assumed that the bias forces in surge and sway, and the yaw moment are slowly varying. A frequently used bias model for marine control applications is the st-order Markow process: _b =?T? b + n (7) where b < is a vector of bias forces and moment, n < is a vector of zero-mean Gaussian white noise, T < is a diagonal matrix of positive bias time constants and < is a diagonal matrix scaling the amplitude 5

LF + WF motion Total motion, LF + WF LF motion WF motion 5 5 time [sec] Fig.. Figure showing the total ship motion as the sum of the LF-motion and the WF-motion. Notice that the st-order WF-motion is osillatoric. of n. This model can be used to describe slowly-varying environmental forces and moments due to (Fossen 994): { nd-order wave drift { ocean currents { wind { unmodeled dynamics.4 st-order Wave-Induced Model A linear wave frequency (WF) model of order p can in general be expressed as: _ = + w (8) w =? (9) where < p, w < and ; and? are constant matrices of appropriate dimensions. The WF response of the ship is generated by using the principle of linear superposition, that is the st-order wave-induced motion w = [x w ; y w ; w ] T is added to the LF motion components of the ship given by = [x; y; ] T : Hence, the total ship motion is the sum of the LF-motion components and the WF-motion components as shown in Figure. Notice the oscillatoric behavior of the wave-induced motion component. It is assumed that the WF model excitations w in (8) are zero-mean Gaussian white noise. For notational simplicity, a nd-order wave model for the st-order waveinduced motion is considered in the remaining of the paper. However, higher order wave transfer function approximations can also be used, such as a 4thorder wave model with 5 parameters, see Grimble et al. (98a) and Fung and 6

Grimble (98), and a 6th-order wave model with 4 parameters (Triantafyllou et al. 98). The main reason for choosing a higher order of the WF model is that a more precise approximation to the actual wave spectrum e.g. the Pierson-Moskowitz or the JONSWAP spectra can be obtained (Fossen 994). However, this also increases the number of model parameters to be determined and the dimension of the observer gain matrices to be tuned. Wave Spectrum Approximation (nd-order Wave-Induced Motion) The nd-order wave-induced motion model was originally proposed by Balchen et al. (976) who used three harmonic oscillators. Slid et al. (98) improved the WF model approximation by including an additional damping term. This model can be written (one model for each degrees of freedom): h i w(s) = i s s + i! i s +! i () where! i (i=...) is the dominating wave frequency, i (i=...) is the relative damping ratio and i (i=...) a parameter related to the wave intensity. A state-space realization of () in degrees of freedom (DOF) is: 64 _ 75 = 64 _ w = where <, < and: I 75 64 75 + 64 75 w () 6 4 75 I n =?diag! ;! ;! =?diag f! ;! ;! g = diag f ; ; g : o.5 Measurement System For conventional ships only position and heading measurements are available. Several position measurement systems are commercially available, such as hydroaccustic positioning reference (HPR) systems and satellite navigation systems. The two commercial available satellite systems are Navstar GPS (USA) and GLONASS (Russian), see Parkinson (996). The accuracy of the GPS satellite navigation system is degraded for civilian users and this results in root-mean-squares errors of about 5 (m) for the horizontal positions. 7

However, for ship positioning systems this problem is usually circumvented by using a dierential global positioning system (DGPS). The main idea of a dierential GPS system is that a xed receiver e.g. located on shore with known position, is used to calculate the GPS position errors. The position errors are then transmitted to the GPS-receiver on board the ship and used as corrections to the actual ship position. In a DGPS-system the horizontal positioning errors are squeezed down to less than (m) which is the typical accuracy of a ship positioning system today (Hofmann-Wellenhof et al. 994). The heading of the vessel is usually measured by using a gyro compass where gyro drift can be compensated for by using a magnetic compass. The accuracy of a gyro compass is typically in the magnitude of. (deg). Hence, the position and heading measurement equation can be written: y = + w +v () where w is the vessel's WF motion due to st-order wave-induced disturbances and v < is zero-mean Gaussian white measurement noise. It is important that the observer is tuned such that the observer gains reect the dierences in the noise levels of the measured signals. In addition to these measurements, the ship observer needs actuator measurements u such that the control forces in surge and sway, and moment in yaw: = B u u () can be computed. Recall that the input matrix B u is assumed to be known. Denition (Wave Filtering). Wave Filtering can be dened as the reconstruction of the LF motion components from the noisy measurement y = + w +v by means of an observer (state estimator). In addition to this, a noise-free estimate of the LF velocity should be produced from y. This is crucial in ship motion control systems since the oscillatoric motion w due to st-order wave-induced disturbances will, if it enters the feedback loop, cause wear and tear of the actuators and increase the fuel consumption. Remark. In general it is impossible to counteract the st-order waveinduced motion of a ship when applying a reasonable propulsion and thruster system. Hence, no improvement in position performance should be expected by feeding back the signal w to the controller. 8

Resulting System Model In the Lyapunov stability analysis, the following assumptions are made: A: M = M T > : This is true if starboard and port symmetries, and low speed are assumed. Symmetry and positive deniteness are only necessary in the Lyapunov analysis. In addition to this, it is assumed that the added mass terms are independent of the wave-frequency such that _M = holds. This is a good assumption for most low-speed applications (Fossen 994). A: n = w = : The bias and wave models are driven by zero-mean Gaussian white noise. These terms are omitted in the observer model and Lyapunov function analysis since the estimator states are driven by the estimation error instead. A: v = : Zero mean Gaussian white measurement noise is not included in the Lyapunov function since this term is negligible compared to the storder wave disturbances w : Full-scale experiments have shown that the observer is highly robust for measurement noise produced by commercial available DGPS-systems and gyro compasses. A4: J() = J(y) or J( ) = J( + w ): This is a good assumption since the magnitude of the wave-induced yaw disturbance w will be less than 5 degrees in extreme weather situations (sea state codes 5{), and less than degree during normal operation of the ship (sea state codes {5). Application of Assumptions A{A4 to (), (), (7), () and () yields the following system model: _ = (4) _ = J(y) (5) _b =?T? b (6) M _ =?D + J T (y)b+ (7) y = + w = +?: (8) For notational simplicity (4), (5) and (8) are written in state-space form: where = [ T ; T ] T and: A = 64 _ = A +B J(y) (9) y = C () 75 ; B = 64 75 ; C =? I : () I 9

4 Nonlinear Observer Design The objective of the observer design is to generate LF position and heading estimates, estimates of the unmeasured LF velocities and estimates of the bias terms by using the idea of wave ltering as stated in Denition. Since the measurements consist of both a LF and a WF component, a wave model is included in the observer to produce a notch eect. When the dominating wave frequency is known, the observer should have the properties of a notch-lter in the frequency range of the wave disturbances. Notice that the notch frequency (dominating wave frequency) will be inside the bandwidth of most ships. 4. Observer Equations A nonlinear observer copying the dynamics (4){(8) is: _^ = ^ + K ~y () _^ = J(y)^ + K ~y () _^b =?T?^b+ ~y (4) M _^ =?D^ + J T (y)^b + + J T (y)k~y (5) ^y = ^ +?^ (6) where ~y = y? ^y is the estimation error and K < 6 ; K < ; K < and < are observer gain matrices to be interpreted later. > is an additional scalar tuning parameter motivated by the Lyapunov analysis, see Section 5.. Similarly as (9) and (), the system (), () and (6) is written in state-space form: _^ = A ^ +B J(y)^ + K ~y (7) ^y = C ^ (8) where ^ = [^ T ; ^ T ] T and: K = 64 K K 75 : (9) The observer structure is shown in Figure.

g K T J (y) Bias estimator: wave drift forces currents - wind T J (y) b^ T - - g L K K W DGPS Compass Wave estimator: - st-order wave-induced motion ^x G y~ - y^ y u B u t Actuators - M - D n^ J(y) h^ ^h w x, ^ y,y ^, ^ u, ^ v, ^ r^ Fig.. Block diagram showing the nonlinear observer. 4. Estimation Error Dynamics The estimation errors are dened as ~ =? ^, ~b = b? ^b and ~ =?^ : Hence, the error dynamics can be written: _~ = (A? KC ) ~ +B J(y)~ () _~b =?T? ~b? ~y () M _~ =?D~ + J T (y)~b? J T (y)k~y: () The dynamics of the velocity estimation error () can be rewritten as: M _~ =?D~? J T (y)~z where ~z, K~y?~b: () By dening ~x, 64 ~ ~b 75 (4)

T J (y) e - z g - M - D n H n J(y) ~ z g b K T - - - g L ~ y - C h ~ A K - B H e n Fig. 4. Block diagram showing the dynamics of the position/bias and velocity estimation errors. (), () and () can be written in compact form as: _~x = A~x + BJ(y)~ (5) ~z = C~x (6) where A = 64 A? KC 75 ; B = 64 B? C?T? 75 ; C = KC?I : (7) The error dynamics is shown in Figure 4 where two new error terms " z =?J T (y)~z and " = J(y)~ are dened.

5 Stability Analysis In this section both GES (Section 5.) and passivity (Section 5.) of the observer are proven. 5. SPR-Lyapunov Function Analysis Lemma (Kalman-Yakubovich-Popov (KYP) Lemma). Let Z(s) = C(sI? A)? B be a n n transfer function matrix, where A is Hurwitz, (A; B) is controllable, and (A; C) is observable. Then, Z(s) is strictly positive real (SPR) if and only if there exist positive denite matrices P = P T and Q = Q T such that (Khalil 996): PA + A T P =?Q (8) B T P = C (9) Proposition (SPR Velocity Estimation Error). If > and the observer gain matrices are given the following structure: K = K = 64 64 k k k k k k k k k 75 ; K = 64 75 ; = 6 4 75 75 (4) (4) then the elements k ij > ; i > and i > can be chosen such that the triple (A,B; C) given by (7), that is the mapping " 7! ~z satises the KYP lemma. z~ H B (s) = h (s) B h B(s) h (s) B y ~ H (s) = h (s) h (s) Fig. 5. The decoupled transfer function structure of the position error dynamics. h (s) e n

Proof. Since K and are chosen to be diagonal, the mapping " 7! ~z (see the lower block in Figure 4) can be described by three decoupled transfer functions: and: ~z(s)= H(s)" (s) where H(s) = H (s)h B (s) H (s) = C (si + A? KC )? B H B (s) = K + (si + T? )? : The diagonal structure of H(s) is shown in Figure 5. The transfer functions h i (s) (i=...) of H (s) and h i B (s) (i=...) of H B(s) of becomes: h i (s) = s + i! oi s +! oi s + (k i + k i + i! oi )s + (! oi + i! oi k i? k i! oi)s +! oi k i (4) h i B (s) = i s + T i s + T i + i i T i i s + i i s + T i : (4) In order to obtain the desired notch eect (wave ltering) of the observer, see Denition, the desired shape of h i (s) is specied as: h i d(s) = s + i! oi s +! oi (s + ni! oi s +! oi) (s +! ci ) (44) where ni > i determines the notch and! ci >! oi is the lter cut-o frequency. Equating (4) and (44) yields the following formulas for the lter gains in K and K : k i =?! ci ( ni? i )! oi (45) k i =! oi ( ni? i ) (46) k i =! ci : (47) Notice that the lter gains can be gain-scheduled with respect to the dominating wave frequencies! oi if desired. In Figure 6 the total transfer function h i (s) = h i (s) B hi (s) is illustrated when all lter gains are properly selected. It is important that the decoupled transfer functions h i (s) all have phase greater than?9 in order to meet the SPR requirement. It turns out that the KYP lemma and therefore the SPR requirement can easily be satised if the following tuning rules for T i ; i and i are applied: =T i i = i <! oi <! ci ; (i=...). (48) 4

- wave drift forces - currents S( w) Wave spectrum w db - - - - - -4 deg - -4-6 -8-9 -4 - - - T i l k i i w oi w ci log w Fig. 6. Bode plot showing the transfer functions h i (s) when =T i i = i <! oi <! ci (i=...). Here! oi (i=...) are the dominating wave frequencies and T i (i=...) are the bias time constants used to specify the limited integral eect in the bias estimator. M Remark. The observer can be gain-scheduled with respect to the dominating wave frequency vector! o = [! o ;! o ;! o ] T by noticing that K = K (! o ); see Eqs. (45){(46). The gain matrices K ; and K are independent of! o : An estimate of! o can be found by using an on-line frequency tracker (Fossen 994). Theorem (Main Result: Globally Exponentially Stable Nonlinear Observer). Under Assumptions A{A4 the nonlinear observer given by (){(6) is globally exponentially stable. Proof. Consider the following Lyapunov function candidate: V = ~ T M ~ + ~x T P ~x: (49) 5

T J (y) e z n - ~ z H H e n J(y) Fig. 7. The observer error dynamics in terms of two interconnected feedback error systems H (velocity) and H (position and bias). Dierentiation of V along the trajectories of ~ and ~x and application of Assumptions A{A4, yields: _V =?~ T D + D T ~ + ~x T P A + A T P ~x +~ T J T (y)b T P ~x?~ T J T (y)~z: (5) Application of Proposition to (5), yields: _V =?~ T D + D T ~? ~xt Q~x: (5) Hence, ~ and ~x = [~ T ; ~ T ; ~b T ] T M converge exponentially to zero, q.e.d. 5. Passivity Interpretation of the Nonlinear Observer The error dynamics in Figure 4 can be described as illustrated in Figure 7 by letting H denote the mapping " z 7! ~ and H denote the mapping " 7! ~z. Notice that the coordinate transformation is performed through a non-singular and bounded matrix J(y). Proposition (Strictly Passive Velocity Error Dynamics). The mapping H is state strictly passive. Proof. Let U = ~T M ~ (5) be a positive denite storage function. Time dierentiation of U along the trajectories of ~ yields: _U =? ~T D + D T ~? ~z T J(y)~ (5) 6

Using the fact that " z =?J T (y)~z; yields: Hence: " T z ~ = _ U + ~T D + D T ~ (54) Z t t " T z ()~() d ~T ~+ (55) where = min(m) is a positive constant and = R t t ~ T D + D T ~d is the dissipated energy due to hydrodynamic damping. M Theorem (Passive Observer). The nonlinear observer (){(6) is passive. Proof. Since it is established that H is state strictly passive and H is SPR (Propositions and ), the nonlinear observer (){(6) must be passive. In addition to this, GES has been shown (Theorem ). M 6 Case Study Both computer simulations and full-scale experiments have been used to evaluate the performance and robustness of the nonlinear passive observer. The case studies are based on the following models of the ship-bias-wave system: M = D = 64 64 5: 6 8:8 6 :7454 9 75 (56) 5:4 4 :79 5?4:9 75 6 : (57)?4:9 6 4:894 8 A picture of the actual ship is shown in Figure 8. The length of Northern Clipper is L = 76: (m) and the mass is m = 4:59 6 (kg). The coordinate system is located in the center of gravity. The bias time 7

Fig. 8. Nortern Clipper. A multipurpose supply vessel owned by Svik Supply Management, Norway. constants were chosen as: T = 64 75 (58) The wave model parameters were chosen as i = : and! oi = :8976 (rad/s) corresponding to a wave period of 7: (s) in surge, sway and yaw. The notch lter parameters were chosen as ni = : and! ci = :: From (45){(47) we get: K = 64?:59?:59?:59 :657 :657 :657 75 ; K = 64 : : : 75 (59) 8

The last two gain matrices and were chosen as: K = 64 : : : ; =:K, = (6) 75 Both the simulation study and the full-scale experiment were performed with a measurement frequency of (Hz). The simulation study was performed with non-zero noise terms n; v and w even though these terms were assumed to be zero in the Lyapunov analysis. This was done to demonstrate the excellent performance of the observer in the presence of stochastic noise. 6. Simulation Study In the simulation study the control inputs were chosen as: = 64 sin(:5t) sin(:t) sin(:7t) 75 (6) The amplitudes of the st-order wave-induced motion in surge, sway and yaw were limited to. (m),. (m) and. (deg), respectively. These disturbances were added to the positions and heading measurements in the simulator. The simulation results are shown in Figures 9{. The st-order wave-induced motion in surge, sway and yaw and their estimates are shown in the lower left plots. It is seen that excellent tracking of the position, velocity and bias is obtained even though the measurements are highly noise-corrupted. In particular, it is impressing that the stochastic behavior of the bias is estimated as well. This means that the zero white noise assumption in the proof can be relaxed when implementing the observer. 6. Experimental Results The simulated observer was implemented on Northern Clipper (Figure 8) to investigate the performance and robustness on "real" data. The time-series from the full-scale experiments are shown in Figures {4. The experiments were performed in an extreme weather situation to demonstrate the ltering properties. The amplitudes of the st-order wave-induced motion in surge, sway and yaw were as large as 8. (m),. (m) and. (deg), respectively. 9

This corresponds to sea state codes 8 and 9 (very high or phenomenal seas). Again, excellent results were obtained. It is seen from the plots that most of the st-order wave disturbances are ltered out resulting in smooth estimates of the positions, heading and velocities. Hence, the feedback controller should not be to much aected by the rough weather condition. 7 Conclusions In this paper a nonlinear passive observer for ships in surge, sway and yaw has been derived. The observer is proven to be global exponential stable and passive. The number of tuning parameters has been reduced to a minimum by using passivity theory. It has been shown that both the low-frequency position and velocity of the ship together with accurate bias state estimates (environmental disturbances) can be computed from noisy position measurements. In addition, ltering of st-order wave-induced disturbances has been done. The nonlinear passive observer has been simulated on a computer model of a supply vessel and implemented on a full-scale supply vessel (Northern Clipper) with excellent results. The simulation results showed that all estimation errors including the bias estimation errors converged exponentially to zero. Experimental results from sea trials performed in the North Sea have also been reported. The main advantage of the nonlinear design to a linear design is that the kinematic equations of motion do not have to be linearized about a set of predened constant yaw angles, typically 6 operating points in steps of degrees, to cover the whole heading envelope of the ship. This is, however, common when using linear Kalman lters and gain scheduling techniques. In addition, it is shown that the tuning procedure is relatively simple to perform due to the passive structure of the observer. On the contrary, it is quite time consuming to tune the Kalman lter since the weighting matrices will be of dimension 5 5 and corresponding to the estimation error and control input vectors, respectively. The results of the paper have applicability to dynamic positioning (DP), tracking control and berthing of ships as well as other marine craft. Future work will look into the design of a nonlinear control law utilizing the passive structure of the observer.

Acknowledgement The authors are grateful to Dr. Asgeir J. Srensen and co-workers at ABB Industri AS in Oslo for valuable comments, discussions and help with the sea trials. References [] Aarset, M. F., J. P. Strand and T. I. Fossen. Nonlinear Vectorial Observer Backstepping With Integral Action and Wave Filtering for Ships, Proceedings of the IFAC Conference on Control Applications in Marine Systems (CAMS'98), Fukuoka, Japan, 7{ October. [] Balchen, J. G., N. A. Jenssen and S. Slid (976). Dynamic Positioning using Kalman Filtering and Optimal Control Theory. Proceedings of the IFAC/IFIP Symposium on Automation in Oshore Oil Field Operation, Holland, Amsterdam, pp. 8{86. [] Balchen, J. G., N. A. Jenssen and S. Slid (98a). Dynamic Positioning of Floating Vessels based on Kalman Filtering and Optimal Control. Proceedings of the 9th IEEE Conference on Decision and Control, New York, NY, pp. 85{864. [4] Balchen, J. G., Jenssen, N. A. and Slid, S. (98b). A Dynamic Positioning System based on Kalman Filtering and Optimal Control. Modeling, Identication and Control, MIC-():5{6. [5] Fossen, T. I. (994). Guidance and Control of Ocean Vehicles (John Wiley & Sons Ltd.). [6] Fossen, T. I., S. I. Sagatun and A. J. Srensen (996). Identication of Dynamically Positioned Ships, Journal of Control Engineering Practice, JCEP- 4():69{76. [7] Fossen, T. I. and A. Grvlen (998). Nonlinear Output Feedback Control of Dynamically Positioned Ships Using Vectorial Observer Backstepping. IEEE Transactions on Control Systems Technology, TCST-6():{8. [8] Fossen, T. I., A. Loria and E. Panteley. Nonlinear Global Output Feedback Control of Ships Using the Separation Principle, Submitted to the IEEE Conference on Decision and Control (CDC'98). [9] Fung, P. T.-K. and M. Grimble (98). Dynamic Ship Positioning using Self-Tuning Kalman Filter. IEEE Transactions on Automatic Control, TAC- 8():9{49. [] Grimble, M. J., R. J. Patton and D. A. Wise (98a). The Design of Dynamic Ship Positioning Control Systems using Stochastic Optimal Control Theory. Optimal Control Applications and Methods, OCAM-:67{.

[] Grimble, M. J., R. J. Patton and D. A. Wise (98b). Use of Kalman Filtering Techniques in Dynamic Ship Positioning Systems. IEE Proceedings Vol. 7, Pt. D, No., pp. 9{. [] Homan-Wellenhof, B., H. Lichtenegger and J. Collins (994). GPS Theory and Practice, Springer-Verlag, New York. [] Khalil, H. K. (996). Nonlinear Systems (Prentice Hall Inc.) [4] Parkinson, B. (ed.) (996). Global Positioning System: Theory and Applications, (American Institute of Aeronautics and Astronautics, Inc., Washington). [5] The Society of Naval Architects and Marine Engineers - SNAME (95). Nomenclature for Treating the Motion of a Submerged Body through a Fluid, Technical and Research Bulletin, No. -5. [6] Slid, S., N. A. Jenssen and J. G. Balchen (98). Design and Analysis of a Dynamic Positioning System based on Kalman Filtering and Optimal Control. IEEE Transactions on Automatic Control, TAC-8():{9. [7] Srensen, A. J., S. I. Sagatun and T. I. Fossen (996). Design of a Dynamic Positioning System using Model-Based Control. Journal of Control Engineering Practice, JCEP-4():59{68. [8] Triantafyllou, M. S., M. Bodson and M. Athans (98). Real Time Estimation of Ship Motions using Kalman Filtering Techniques. IEEE Journal of Oceanic Engineering, JOE-8():{4.

actual and estimated x [m] actual and estimated u [m/s].5 -.5 - - 4 -.5 4 actual and estimated xw [m] actual and estimated b [kn] - -4 - -6-4 time [sec] -8 4 time [sec] Fig. 9. Simulation study: Position, velocity, wave disturbance and bias (grey), and their estimates (black) in surge.

5 actual and estimated y [m]. actual and estimated v [m/s] -5 - -. -5 -. - 4 -. 4 actual and estimated yw [m] actual and estimated b [kn].5 - -4 -.5-6 - 4 time [sec] -8 4 time [sec] Fig.. Simulation study: Position, velocity, wave disturbance and bias (grey), and their estimates (black) in sway. 5 actual and estimated psi [deg]. actual and estimated r [deg/s] -5 - -. -5-4 -. 4 actual and estimated psiw [deg] actual and estimated b [knm] 5 - -5-4 time [sec] - 4 time [sec] Fig.. Simulation study: Heading angle, angular rate, wave disturbance and bias (grey), and their estimates (black) in sway. 4

4 tau and tau [kn] - 4 5 6 7 8 5 tau [knm] -5-4 5 6 7 8 time [sec] Fig.. Experimental data: Control inputs in surge, sway and yaw. GPS-meas. and estimated x [m] estimated xw [m] 5-5 - -5 4 6 8 5-5 - 4 6 8 GPS-meas. and estimated y [m] estimated yw [m] -5 - -5 - -5 4 6 8 time [sec] 5-5 - 4 6 8 time [sec] Fig.. Experimental data: Actual LF+WF position (grey) with estimates (black) of the LF- and WF-position components in surge and sway. 5

Gyro-meas. and estimated psi [deg] 75 7 65 6 55 - estimated psiw [deg] 5 4 6 8-4 6 8 estimated u and v [m/s]. estimated r [deg/s].5 -.5 4 6 8 time [sec].5 -.5 -. 4 6 8 time [sec] Fig. 4. Experimental data: Upper plots { Actual LF+WF heading angle (grey) with estimates (black) of the LF- and WF-heading angles. Lower plots { Estimates of the LF velocities in surge, sway and yaw. 6